Vision: [Legal question] Is it legal to use pre-trained models for commercial purposes?

[ternaus]

[ternaus]

I see discussions at

• Reddit
• Kaggle
• PyTorch forum
• Authors Guild, Inc. v. Google, Inc.

But there is no definitive answer.

[ternaus]

Similar question

• to Google https://github.com/tensorflow/models/issues/9131
• to Intel https://github.com/openvinotoolkit/open_model_zoo/issues/1433

[fmassa]

I'll leave this to @dongreenberg to answer

[dongreenberg]

Hi @ternaus - sorry for the delay on this. The short answer is that licenses for the underlying data are generally pretty generous about this type of derivative use, but they occasionally vary. Our legal team have said that this is entirely a matter of your own or your legal team's comfort with the use you have in mind and the underlying license, so there really isn't a boxed answer here unfortunately. I'd encourage you to check with your legal teams to see what they're comfortable with. For onlookers without large legal teams, one option may be to ask the data publisher directly.

[ternaus]

@dongreenberg Thank you for looking at this issue.

licenses for the underlying data are generally pretty generous about this type of derivative use

Not really.

Most of the models at https://pytorch.org/docs/stable/torchvision/models.html are trained on ImageNet that was scrapped from the internet.

For the database of the image URLs: Researcher shall use the Database only for non-commercial research and educational purposes.

In addition, Images at these URLs could be subject to copyright.

one option may be to ask the data publisher directly.

Correct me, if I am wrong, but this means reaching out to millions of people that took pictures in the ImageNet dataset. Researchers that collected a set of URLs to these images do not have any rights for these photos.

I understand that legal in Machine Learning is a wild west and things like this is really a gray area.

I am confident that many companies do not even know that using pre-trained models that were trained on non-commercial datasets could be an issue. To increase the transparency and raise awareness, is it possible that you add words like:

It is entirely a matter of your own or your legal team's comfort to use these pre-trained models in commercial applications.

to the top of https://pytorch.org/docs/stable/torchvision/models.html and https://pytorch.org/hub/?

[dongreenberg]

Sorry, to be clear:

1. I know absolutely nothing about this, just relaying our legal team's guidance.
2. The license for use of the _datasets_ may not be generous, but the "generosity" I've been told is in the interpretation that weights of a model trained on that data may be considered derivative enough to be ok for commercial use. Again, this is a subjective matter of person comfort. There is no publishable "answer" we can give.
3. Absolutely, that's a great suggestion. I will have to double check that that's alright to put out there, again because my personal sense of law is comparable to my personal sense of heart surgery.